RESEARCH. Noa Dagan, 1,2 Chandra Cohen-Stavi, 1 Maya Leventer-Roberts, 1,3 Ran D Balicer 1,4. open access

Transcription

1 open access External validation and comparison of three prediction tools for risk of osteoporotic fractures using data from population based electronic health records: retrospective cohort study Noa Dagan, 1,2 Chandra Cohen-Stavi, 1 Maya Leventer-Roberts, 1,3 Ran D Balicer 1,4 1 Clalit Research Institute, Chief Physician s Office, Clalit Health Services, Tel Aviv, Israel 2 Computer Science Department, Ben Gurion University of the Negev, Be er Sheba, Israel 3 Department of Preventive Medicine and Department of Pediatrics, Icahn School of Medicine at Mount Sinai, New York, New York, USA 4 Epidemiology Department, Ben Gurion University of the Negev, Be er Sheba, Israel Correspondence to: N Dagan noa.dgn@gmail.com Additional material is published online only. To view please visit the journal online. Cite this as: BMJ 2017;356:i Accepted: 08 December 2016 ABSTRACT Objective To directly compare the performance and externally validate the three most studied prediction tools for osteoporotic fractures QFracture, FRAX, and Garvan using data from electronic health records. Design Retrospective cohort study. Setting Payer provider healthcare organisation in Israel. Participants members aged 50 to 90 years for comparison between tools and cohorts of different age ranges, corresponding to those in each tools development study, for tool specific external validation. Main outcome measure First diagnosis of a major osteoporotic fracture (for QFracture and FRAX tools) and hip fractures (for all three tools) recorded in electronic health records from 2010 to Observed fracture rates were compared to probabilities predicted retrospectively as of Results The observed five year hip fracture rate was 2.7% and the rate for major osteoporotic fractures was 7.7%. The areas under the receiver operating curve (AUC) for hip fracture prediction were 82.7% for QFracture, 81.5% for FRAX, and 77.8% for Garvan. For major osteoporotic fractures, AUCs were 71.2% for QFracture and 71.4% for FRAX. All the tools underestimated the fracture risk, but the average observed to predicted ratios and the What is already known on this topic Tools for prediction of osteoporotic fractures are recognised by leading guidelines as an important component of osteoporosis prevention but are underutilised Of the three most studied fracture prediction tools QFracture, FRAX, and Garvan QFracture was the only one developed using data from electronic health records The adaptation of these tools for automatic implementation in external electronic health record systems is not clear, nor is their relative performance What this study adds Automatic computation of all three tools using data from external electronic health records produced similar results, as has been previously reported, for each of the tools separately (tested in separate cohorts of the same age ranges as the derivation cohort of each tool) When evaluated using one cohort (for which the age ranges of all tools overlap), QFracture and FRAX yielded high discriminatory performance for hip fracture prediction, with QFracture performing slightly better This performance gap was much smaller than previously reported by reviews, which compared results from validation studies that tested each of the tools using different age ranges calibration slopes of FRAX were closest to 1. Tool specific validation analyses yielded hip fracture prediction AUCs of 88.0% for QFracture (among those aged years), 81.5% for FRAX (50-90 years), and 71.2% for Garvan (60-95 years). Conclusions Both QFracture and FRAX had high discriminatory power for hip fracture prediction, with QFracture performing slightly better. This performance gap was more pronounced in previous studies, likely because of broader age inclusion criteria for QFracture validations. The simpler FRAX performed almost as well as QFracture for hip fracture prediction, and may have advantages if some of the input data required for QFracture are not available. However, both tools require calibration before implementation. Introduction Osteoporotic fractures cause major morbidity and mortality, with many people who experience such fractures rapidly deteriorating in health status and experiencing a lower quality of life. 1 2 This poses a substantial economic burden to health systems, patients, and their families. 3 The burden of osteoporotic fractures is expected to increase as populations age, with the incidence of hip fractures reported to increase 30-fold between the ages of 50 and 90 years. 4 Osteoporotic fractures and re-fractures can be prevented and better managed when people at high risk are identified early. 5 6 Routine scanning of bone mineral density is recommended in all women and in some guidelines also for men, but despite these recommendations, rates of screening remain low, leaving osteoporosis undiagnosed in many patients Furthermore, the criteria used for bone mineral density to identify those at high risk for osteoporotic fractures are not highly sensitive, as more than half of older women with osteoporotic fractures do not meet the bone mineral density criteria for osteoporosis (T score lower than 2.5). 11 For these reasons, multiple risk assessment tools based on clinical and personal characteristics have been developed in recent years to identify those at high risk for osteoporotic fractures. The most studied tools are the World Health Organization s FRAX, Garvan, and QFracture, which are all freely available online for public use. 12 Each tool has been developed in different contexts, with FRAX and Garvan based on cohort studies using survey and doctor and patient reported data, and QFracture based on data from electronic health records The extent to which each tool has been externally validated varies: FRAX has been validated by 26 studies in nine countries, Garvan by six studies in the bmj BMJ 2017;356:i6755 doi: /bmj.i6755 1

2 three countries, and QFracture by three studies within the United Kingdom and the Republic of Ireland. 12 These three tools also differ in their complexity in terms of the number of input variables included, with QFracture using 26 variables, FRAX using 11, and Garvan using five. In addition, FRAX and Garvan offer predictions with or without the input of a pre-existing bone mineral density measurement, whereas QFracture does not include bone mineral density in its algorithm. Supplement 1 summarises the basic features of the three tools. Although many predictive tools have been developed, few are used to support clinical decision making to identify patients at high risk for osteoporotic fractures. 17 With increasing use of electronic health record systems there is the potential to produce automated personalised fracture risk scores to better direct treatment and reduce the overall burden of osteoporotic fractures. These risk scores can be presented both directly to the patients and made accessible to their doctors through the electronic health record system. Several studies have shown the benefits of improved management of osteoporosis and fracture prevention from electronic health records or electronic software based decision support implementations In determining which of the various prediction tools is adaptable for automatic implementation using electronic health record data, the predictive performance of each tool (both discrimination and calibration), the validation results in various populations, and the availability of types of data required for the tool must be considered. Although numerous reviews have compared FRAX, Garvan, and QFracture, to the best of our knowledge their performance has not been directly compared within one population. A few studies have directly compared two of the tools in the same population However, the only study to evaluate tool performance in a large population and among both men and women compared old versions of QFracture and FRAX. 13 Several other studies purported to compare two tools but did not validate the predicted risk with observed events over a subsequent follow-up period Several substantial pitfalls have been highlighted both in comparisons of performance across various tools and in external validations of specific tools. Problems with missing input variables, sample size, and the number of outcome events were noted as limiting the ability of validation studies to provide generalisable results and full validations of original tools. 17 Most studies aiming to compare measures of tool performance were reviews or meta-analyses (not direct comparisons within one population) that relied on results of specific tool validations. The comparability of these validations has been critiqued, because different inclusion criteria and follow-up periods might affect their results. 30 Age, for example, is a major determinant of fracture risk, and thus the choice of age ranges included in specific validation studies were suggested to substantially affect the reported performance of the tool. Furthermore, many of these validations did not present a comparison between the validation and derivation populations used to develop the tools, to shed light on the kind of validation they contribute ranging between reproducibility (evaluating the tool within a population with similar characteristics) and transportability (evaluating the tool within a population of different characteristics). 31 The lack of consistency in study designs among previous validation studies presents challenges in the ability to draw meaningful conclusions about which tool offers the best performance. We compared the performance of the three most commonly studied fracture prediction tools in a single, large population when computed automatically based on electronic health record data. We also conducted a tool specific external validation in an independent population to evaluate the performance of the tools in populations with the same age range as those in which they were developed, thus allowing comparison with previously reported performance. Methods Setting This study used electronic health record data from Clalit Health Services, the largest of four national health funds in Israel. All Israeli residents are covered by one of the health funds and can switch between them at any time; however, switching rates are relatively low about 1% annually 32 which allows for consistent longitudinal follow-up. Clalit Health Services is both a healthcare insurer and a provider, thus financing and supplying services to its 4.3 million members, which make up more than half of Israel s population. Membership of Clalit Health Services comprises the general population, but for historical reasons the organisation has a slightly larger proportion of the older population and those from a lower socioeconomic class. 33 Study design In this historical prospective cohort study we compared the probability of hip fracture over five years using FRAX, QFracture, and Garvan, as well as the probability of major osteoporotic fractures over five years using FRAX and QFracture, computed on 1 January 2010 (index date), with fracture events observed up to 31 December 2014 (follow-up period). In the first part of this study we compared the performance of the three tools, and thus selected a population in which the reported age ranges for all tools overlap. This comparative analysis was conducted for risk of hip fracture. Because the definition used by Garvan for major osteoporotic fractures is much broader than the one used by QFracture and FRAX (vertebral, distal radius, proximal humerus, or hip), we conducted additional analyses only between QFracture and FRAX to compare the performance for predicting major osteoporotic fractures. In the second part of the study we conducted a tool specific external validation for performance in predicting fractures, using cohorts with varying age ranges for each tool. Study population The comparative analysis was performed among members of Clalit Health Services aged 50 to 90 years as of 2 doi: /bmj.i6755 BMJ 2017;356:i6755 the bmj

3 the bmj BMJ 2017;356:i6755 doi: /bmj.i6755 the index date, who had at least three years of continuous membership before the index date and through the follow-up period or until death (see fig 1). Therefore, the cohort did not include those who were lost to follow-up. Although FRAX was developed among a population that excluded patients who were treated for osteoporosis, 34 the other two tools were not, and thus for comparative purposes, treatment for osteoporosis was not used as an exclusion criterion (a population of non-treated patients was evaluated in a separate sensitivity analysis). For the tool specific external validation analyses we used specified age ranges corresponding to those chosen in the original tool development studies: the QFracture analysis included members aged years, 16 the FRAX analysis included members aged years, 4 and the Garvan analysis included members aged years 14 (the official calculator computes risk for 50 or more years, which was the reason why we chose age 50 as the lower limit for the comparative analysis). 35 The rest of the inclusion and exclusion criteria did not differ from those used in the comparative analysis (see fig 1). To account for real world settings, in the populations of both analyses we included those who died during the follow-up period. Data sources The electronic health record data at Clalit Health Services contain comprehensive administrative and clinical data. These include demographic information, diagnoses given in a community or a hospital setting, chronic disease and oncology registries, laboratory results, written prescriptions and prescriptions dispensed, clinical markers (eg, body mass index, smoking status), medical procedures, and imaging data. Input variables Input variables included clinical status, prescription drug use, and demographic characteristics, according to the variables used in each of the tools. Supplement 2 lists the codes used to define diagnoses and drug based variables. To provide as comprehensive data as possible for the prediction tools, we based all input variables of the three prediction tools on information that was last documented as of the index date. Most study variables represent chronic conditions and were consequently taken with no date limitation before the index date. For variables that could potentially change over time (including body mass index, smoking status, alcoholism, nursing home residency, history of falls, and drug use), we took the last relevant documented history with no time limitation, and we also conducted a sensitivity analysis in which the extraction of such variables was limited to two years before the index date. The sensitivity analysis was performed to establish the implications of not limiting the time from which variable data were taken. Clinical diagnoses Input variables for diagnosis included history of osteoporotic fractures, secondary osteoporosis, dementia, Parkinson s disease, epilepsy, diabetes and other endocrine conditions, obstructive airways disease, cardiovascular disease, malabsorption, chronic liver disease, chronic kidney disease, rheumatoid arthritis, systemic lupus erythematosus, and documented history of falls. We extracted these diagnoses from community and hospital records, as well as from the Clalit Health Services chronic disease registry, when appropriate. Diagnoses were defined based on the International Classification of Diseases, ninth revision (ICD-9), International Classification of Primary Care (ICPC), and chronic disease registry codes. Diagnoses made in the community setting were further validated based on doctors accompanying free text diagnosis description, available only in the community records. Body mass index This variable was computed from documented weight and height measurements. Smoking status In the Clalit Health Services database, smoking status is defined as non-smokers, former smokers, or current smokers. In QFracture, three current smoking categories are provided according to the number of cigarettes smoked daily. 36 To avoid the bias of categorising patients in one of the outlying categories, we assigned Clalit Health Services current smokers to the middle category (10-19 cigarettes daily). For FRAX s two category smoking status, we assigned former smokers in our population to the non-smokers category, as was done in the cohorts used to develop FRAX Alcohol consumption The Clalit Health Services database does not include information on alcohol intake, so we defined alcohol consumption as a dichotomous (yes or no) variable, based on diagnoses of alcoholism or alcohol induced chronic complications (ICD-10 codes for related psychiatric diagnoses were used for alcoholism in addition to the ICD-9 and ICPC codes). Of the five alcohol consumption categories provided by QFracture, we assigned individuals with alcohol related diagnoses to the fourth level category (7-9 units daily, using the UK s definition of alcohol unit), since the lower categories were unlikely to cause alcohol related complications, and the highest category might overestimate the alcohol consumption for some of the relevant population. Given the inability to distribute individuals without alcohol related diagnoses to the various alcohol consumption levels, we assigned them to the none (ie, no alcohol intake) category. Family history of fractures A family history of osteoporosis and hip fractures was defined by diagnosis codes indicating such a history and by searching the medical records of the parents of study members, when the family connection was defined within the electronic health record and either parent was a member of Clalit Health Services. Medication use We computed variables for medication use, considered only in QFracture and FRAX, based on pharmacy dispensing data. Glucocorticoid use was defined differently by these tools two prescription records in the last six months by QFracture versus current or past use for more than three months by FRAX. We therefore computed glucocorticoid use as two separate variables. Purchases of antidepressants and 3

4 hormone replacement therapy medications were included only in the QFracture analyses. Nursing home care We considered an individual to be a nursing home resident when the patient s primary clinic or treating doctor were administratively defined as institutional positions. In cases where there was no documentation of body mass index, weight, or smoking status before the index date (the only variables for which missing data could be identified), we used multiple imputation to complete these values. We also performed a complete case sensitivity analysis without imputed variables. Outcome variables Outcome variables included both hip fracture and major osteoporotic fractures, which were defined as fractures of the hip, vertebrae, distal radius, or proximal humerus. These variables were defined based on the records for clinical diagnoses. Predictive tool risk computation We computed the five year risk according to QFracture (2012 version) and Garvan based on their full tool equations To ensure correct automation, we manually validated a few dozen cases against the official calculator sites. Since the current FRAX equations are not published by the authors, we used the FRAX 10 year probability charts calibrated for Israel, stratified by sex, age, body mass index, and number of clinical risk factors, as supplied by the official FRAX site. 37 We multiplied the 10 year probabilities by 0.5 to convert to five year probabilities. The justification for this transformation was established by examining the rate of osteoporotic fracture events over a 10 year period, between 2005 and 2014 (see supplement 3 for further details). All tools were computed without the input of bone mineral density because QFracture does not include this variable and data on bone mineral density were limited in the electronic health record system for the study years. Statistical analysis To compare across the three tools, which were developed using different modelling methods, we used the provided risk probabilities for each tool respectively and treated the outcome as if it were a binary variable (fracture or no fracture). This decision was also guided by the clinical application of these risk predictions tools that doctors and patients perceive the output as risk for the relevant follow-up period, regardless of the methods used to produce it. The closed cohort design facilitated this strategy of treating the outcome as a binary variable, because there was a known outcome for all study members in a fixed follow-up period of five years Since it is clinically important to test the accuracy of the predicted probability of fracture both for people who survive the follow-up period and for those with shorter life spans, we did not account for shortening of the follow-up period due to death. To evaluate the overall ability of each tool to discriminate between those at low risk and those at high risk we used the area under the receiver operating curve (AUC) in both the comparative and the tool specific external validation analyses. We calculated other discriminatory measures sensitivity, specificity, positive and negative predictive values, accuracy, and error for the top 10% and 20% highest risk cut-offs of each tool. In three separate sensitivity analyses we further evaluated the discrimination measures in the comparative analysis: limitation of the time range of variables with less chronic nature, complete case analysis, and a subpopulation that excluded patients who were being treated for osteoporosis in the two years before the index date. Since the AUC is considered a somewhat crude overall discriminatory measure, that might overlook the contribution of specific risk factors that are not prevalent in the population but are potentially clinically significant for an individual patient s risk prediction, 30 we conducted a reclassification analysis between the two tools with the highest AUCs in the comparative analysis. We report the total numbers of patients classified as low risk and high risk using a top 10% cut-off level for the two tools, as well as measures of net reclassification index analysis. 42 The net reclassification index for events is the rate of events that were correctly reclassified as high risk by the tested tool (usually the tool that incorporates more risk factors) minus the rate of events wrongly reclassified as low risk. The net reclassification index for non-events is the parallel measure, and is the rate of non-events that were correctly reclassified as low risk minus the rate of non-events that were wrongly reclassified as high risk. The overall net reclassification index is the combination of net reclassification index for events and net reclassification index for non-events, whereas the more intuitive weighted net reclassification index is the combination of the same values weighted by the relative size of the groups they represent. 43 We calculated standard errors for all net reclassification index values. 44 We assessed the calibration of each tool by comparing the average predicted risk with the observed percentage of those who experienced fractures over the follow-up period, stratified by age groups and separately by 10ths of fracture risk. To provide calibration measures that are not based on grouping of individuals into strata, we compiled calibration aparametric curves, calibration slopes, and calibration-in-the-large values 45 using functions by Harrell et al 46 and added these to calibration plots. Multiple imputation was conducted using 10 iterations and 20 multiple imputations, thus creating 20 full datasets, using functions by Van Buuren et al. 47 We performed all analyses separately on each of these imputed datasets and averaged these to determine the final performance measures. A 95% confidence interval for AUC measures of specific prediction tools as well as for the differences between tools was calculated using Rubin s rules for variance estimation in multiple imputed datasets (by taking into account both the AUC variance of 1000 bootstrap samples within each imputed dataset and the variance of the 20 average AUCs between the imputed datasets). Owing to the 4 doi: /bmj.i6755 BMJ 2017;356:i6755 the bmj

5 the bmj BMJ 2017;356:i6755 doi: /bmj.i6755 nature of the net reclassification index analysis, this analysis was only based on one random imputed dataset. Plots were created using a combined dataset that included all of the separate imputed datasets. All analyses were conducted using R, CRAN version (mice, 47 ROCR, 49 and rms 46 packages). Patient involvement No patients were involved in setting the research question or the outcome measures, nor were they involved in developing plans for design or implementation of the study. No patients were asked to advise on interpretation or writing up of results. There are no plans to disseminate the results of the research to study participants or the relevant patient community. Results As of 1 January 2010, members of Clalit Health Services were aged 50 to 90 years. Of those, we excluded (2.8%) because they did not meet the criteria for continuous membership (fig 1 ). The final population for the comparative analysis consisted of people (54.6% women). This population included (2.7%) who experienced a hip fracture and (7.7%) who experienced a major osteoporotic fracture during the follow-up period (table 1 ). Supplement file 4 provides specific fracture rates stratified by age and sex. A total of (10.8%) people died during the follow-up period. The average length of follow-up was 4.73 years, with total person years of follow-up. Overall, (5.2%) of the records were imputed for weight and body mass index values, and (3.3%) were imputed for smoking status. Table 1 lists the Comparative analysis flow chart All tools Clalit Health Services members aged years (n= ) Non-continuous members for 3 years before index date or during follow-up period (n=30 289) Continuous members aged years (n= ) Tool specific external validation analysis flow chart QFracture Clalit Health Services members aged years (n= ) Non-continuous members for 3 years before index date or during follow-up period (n= ) Continuous members aged years (n= ) Garvan Clalit Health Services members aged years (n= ) Non-continuous members for 3 years before index date or during follow-up period (n=15 329) Continuous members aged years (n= ) Fig 1 Population flowchart for comparative and tool specific external validation analyses (FRAX external validation population is same as population used for comparative analysis) characteristics of the study population by input variables of the three prediction tools, the outcome fracture rates, and which variables were included in each tool. In examining the comparative performance across the tools, QFracture had the highest AUC for hip fracture prediction (82.7%, 95% confidence interval 82.4% to 82.9%), followed closely by FRAX (81.5%, 81.3% to 81.7%). Garvan s AUC for hip fracture prediction (77.8%, 77.5% to 78.1%) was lower (table 2 ). The confidence interval for the difference between the QFracture and FRAX AUCs was %, whereas the confidence interval for the difference between the QFracture and Garvan AUCs was %. Among the highest 10% risk for hip fracture, as predicted in 2010, QFracture identified 45.1% (sensitivity) of those who went on to experience a hip fracture, FRAX 43.6%, and Garvan 36.9%. By targeting those in the 20% highest risk for hip fracture, QFracture identified 68.9% of hip fractures, FRAX 65.8%, and Garvan 57.1%. The specificity and negative predictive values were high and comparable for all three tools (table 2). The QFracture and FRAX discriminatory measures for prediction of major osteoporotic fractures were lower than those for hip fracture prediction. AUCs for both tools were close (QFracture: 71.2%, 71.0% to 71.4%; FRAX: 71.4%, 71.2% to 71.6%, table 2 ). The confidence interval for the difference between the FRAX and QFracture AUCs was %. The sensitivity for the top 10% highest risk group was 26.7% for QFracture, compared with 29.0% for FRAX, and the positive predictive value was 20.7% for QFracture and 22.4% for FRAX. Figure 2 presents the comparisons of the receiver operating curves for all three tools in predicting hip fractures and for QFracture and FRAX in predicting major osteoporotic fractures. The results from the three sensitivity analyses were consistent on the relative performance of the tools in their discriminatory measures to that of the main analysis. Analyses limiting variable data collection to the two years before the index date can be found in supplement 5. Complete case analyses are in supplement 6, and analyses of non-treated patients in supplement 7. We conducted a reclassification analysis between QFracture and FRAX (the two tools that yielded the highest AUCs) to compare how these tools categorised patients into low risk and high risk groups. QFracture, which incorporates more risk factors than FRAX in its prediction model, was considered the reclassifying model in the analysis, so that we could evaluate the prediction increment offered by its added risk factors (table 3). The net proportion of patients who experienced a hip fracture and were correctly reclassified as high risk by QFracture compared with FRAX was 1.50% (net reclassification index for events). The net proportion of patients who experienced a major osteoporotic fracture and were correctly reclassified as high risk by QFracture was 2.31% (net reclassification index for events). For both types of outcomes, the change in the correct reclassification of non-events was less than 0.2%. The net changes in the proportion of patients assigned a more appropriate risk category for prediction of hip 5

6 Table 1 Characteristics of comparative analysis population, by each of the input variables included in QFracture, FRAX, and Garvan Input variables* No (%) in study population Overall (100) (7.7) (2.7) Age group (years): (38.0) (3.8) 1994 (0.5) (28.4) (5.9) 3689 (1.2) (21.1) (11.6) 9465 (4.3) (12.5) (17.2) (9.8) Sex: Men (45.4) (4.9) 8996 (1.9) Women (54.6) (10.1) (3.3) Ethnicity: Black African (1.2) 540 (4.2) 135 (1.1) White (98.8) (7.8) (2.7) Nursing home residency: No (98.7) (7.7) (2.6) Yes (1.3) 1409 (10.6) 900 (6.8) Body mass index category: Obese (29.3) (7.9) 6833 (2.2) Overweight (38.4) (7.5) (2.5) Normal (26.2) (8.4) 9467 (3.4) Underweight 9216 (0.9) 1113 (12.1) 601 (6.5) Missing (5.2) 2548 (4.6) 1089 (2.0) Smoking category: Non-smoker (64.6) (8.5) (2.9) Former smoker (15.5) (7.0) 3901 (2.4) Current smoker (16.6) (5.8) 3130 (1.8) Missing (3.3) 2107 (6.0) 1027 (2.9) Alcoholism: No (98.9) (7.7) (2.6) Yes (1.1) 1077 (9.6) 468 (4.2) Parental hip fracture: No (98.2) (7.8) (2.7) Yes (1.8) 866 (4.6) 142 (0.8) Parental osteoporotic fracture: No (93.9) (8.0) (2.8) Yes (6.1) 2538 (3.9) 315 (0.5) Major osteoporotic fracture: No (93.3) (6.2) (2.0) Yes (6.7) (29.7) 8883 (12.6) No of fractures after age 50 years: (85.2) (5.7) (1.8) (11.3) (16.6) 7408 (6.2) (2.6) 7307 (26.9) 3111 (11.4) (0.9) 3594 (36.5) 1640 (16.7) History of a fall: No (93.9) (6.9) (2.2) Yes (6.1) (20.6) 6685 (10.4) No of falls in past year: (97.8) (7.4) (2.5) (0.9) 2292 (23.5) 1214 (12.5) (0.9) 2272 (22.8) 1106 (11.1) (0.3) 918 (25.2) 451 (12.4) Secondary osteoporosis : No (93.3) (7.5) (2.6) Yes (6.7) 7305 (10.3) 2683 (3.8) Dementia: No (97.7) (7.6) (2.5) Yes (2.3) 3545 (14.7) 2206 (9.2) Parkinson s disease: No (97.9) (7.6) (2.5) Yes (2.1) 3242 (14.3) 1786 (7.9) Major osteoporotic fracture Hip fracture QFracture FRAX Garvan V V V V V V V V V V V V V V V V V V V V (Continued) 6 doi: /bmj.i6755 BMJ 2017;356:i6755 the bmj

7 Table 1 Characteristics of comparative analysis population, by each of the input variables included in QFracture, FRAX, and Garvan Input variables* No (%) in study population Major osteoporotic fracture Hip fracture QFracture FRAX Garvan Epilepsy: No (94.5) (7.5) (2.6) Yes (5.5) 6892 (11.8) 2641 (4.5) Type 1 diabetes: No (99.9) (7.7) (2.7) Yes 1024 (0.1) 116 (11.3) 45 (4.4) Type 2 diabetes: No (72.6) (7.2) (2.3) Yes (27.4) (9.2) (3.7) Other endocrine disorders: No (95.4) (7.6) (2.6) Yes (4.6) 5409 (11.0) 2031 (4.1) Cancer history: No (86.6) (7.3) (2.4) Yes (13.4) (10.6) 5985 (4.2) Obstructive airways disease: No (84.8) (7.3) (2.5) Yes (15.2) (10.2) 5942 (3.7) Cardiovascular disease: No (71.7) (6.6) (1.9) Yes (28.3) (10.6) (4.6) Malabsorption: No (98.9) (7.7) (2.7) Yes (1.1) 1209 (10.1) 420 (3.5) Chronic liver disease: No (98.0) (7.7) (2.6) Yes (2.0) 2295 (10.8) 841 (3.9) Chronic renal disease: No (92.1) (7.4) (2.4) Yes (7.9) 9468 (11.4) 4695 (5.7) Rheumatoid arthritis: No (97.5) (7.6) (2.6) Yes (2.5) 3519 (13.4) 1263 (4.8) V V Systemic lupus erythematosus: No (99.8) (7.7) (2.7) Yes 1980 (0.2) 268 (13.5) 88 (4.4) Drug purchases : Glucocorticoids: No (97.4) (7.6) (2.6) Yes (2.6) 3971 (14.5) 1346 (4.9) V V Antidepressants: No (90.2) (7.1) (2.4) Yes (9.8) (13.3) 5740 (5.5) Hormone replacement therapy: Yes 8663 (0.8) 416 (4.8) 70 (0.8) No (99.2) (7.8) (2.7) V=variables used as input information for specified tool. *Values within each input variable are sorted by predicted fracture rate ie, variable s value that has lowest predicted risk (as defended by prediction tools) appears first. Fracture rate during follow-up period ( ), within population of each subgroup. Garvan uses a weight instead of body mass index. Defined by any of following: type 1 diabetes, osteogenesis imperfecta, hyperthyroidism, hypogonadism, premature menopause, malabsorption, and chronic liver disease. Numbers were calculated using QFracture s definition of drug purchase at least two purchase records in six months before index date. In the case of glucocorticoid use, which is also used by FRAX, the calculation is based on a history of at least 90 days of use (extracted by number of days covered by past purchase records) and resultant numbers, which were similar to those of the QFracture variable, are not presented. fracture and major osteoporotic fracture by QFracture were 0.08% and 0.36%, respectively. Table 4 presents the absolute probabilities of hip fracture that were calculated by each of the three tools, and the calibration of these probabilities with the absolute fracture rates that were observed over the five year follow-up period, by sex and age groups. A majority of the observed-to-predicted ratios for hip fractures were greater than 1, indicating underestimation of the risk by all three tools for both men and women and in almost all age groups. The QFracture and Garvan ratios presented a consistent downward trend with the increase in age groups but were steadier across the different age groups for FRAX. The risk underestimation was most prominent for women in Garvan. In addition, Garvan was the only tool to assign lower mean predicted the bmj BMJ 2017;356:i6755 doi: /bmj.i6755 7

8 Table 2 Comparison of discriminatory measures between QFracture, FRAX, and Garvan of top 10% and 20% high risk score cut-offs by each tool. Values are percentages unless stated otherwise Discriminatory measures* Denominator QFracture FRAX Garvan Top 10% risk Top 20% risk Measure for top 10% risk Measure for top 20% risk Measure for top 10% risk Measure for top 20% risk Measure for top 10% risk Measure for top 20% risk Hip fractures: AUC NA NA Absolute risk cut-off NA NA Sensitivity ( ) 68.9 ( ) 43.6 ( ) 65.7 ( ) 36.9 ( ) 57.1 ( ) Specificity ( ) 81.3 ( ) 90.9 ( ) 81.3 ( ) 90.7 ( ) 81.0 ( ) PPV ( ) 9.2 ( ) 11.6 ( ) 8.8 ( ) 9.8 ( ) 7.6 ( ) NPV ( ) 99.0 ( ) 98.3 ( ) 98.9 ( ) 98.1 ( ) 98.6 ( ) Accuracy ( ) 81.0 ( ) 89.7 ( ) 80.8 ( ) 89.3 ( ) 80.4 ( ) Error ( ) 19.0 ( ) 10.3 ( ) 19.2 ( ) 10.7 ( ) 19.6 ( ) Major osteoporotic fractures AUC NA NA Absolute risk cut-off NA NA Sensitivity ( ) 46.4 ( ) 29.0 ( ) 47.1 ( ) Specificity ( ) 82.2 ( ) 91.6 ( ) 82.3 ( ) PPV ( ) 18.0 ( ) 22.4 ( ) 18.2 ( ) NPV ( ) 94.8 ( ) 93.9 ( ) 94.9 ( ) Accuracy ( ) 79.4 ( ) 86.7 ( ) 79.6 ( ) Error ( ) 20.6 ( ) 13.3 ( ) 20.4 ( ) NA=not applicable; AUC=area under receiver operating characteristic curve (C statistic); PPV=positive predictive value; NPV=negative predictive value. Analyses comparing performance for predicting major osteoporotic fractures were conducted only between QFracture and FRAX because Garvan s definition for major osteoporotic fractures is much broader than either tool. Numbers in parentheses are numerators for measure; numbers contain a decimal component because they are averaged between imputed datasets. *Assessed with five years of follow-up. probabilities for women compared with men in the same age groups. The observed-to-predicted ratios by 10ths of risk and sex were also more consistent for FRAX compared with QFracture and Garvan, which presented declining ratios as risk increased (table 5 ). Figure 3 presents a calibration plot, presenting the observed and predicted rates for each 10th of risk, along with aparametric calibration curves, calibration slopes, and calibration-in-the-large values. The tool specific external validation analyses consisted of three different cohorts (fig 1 ): the FRAX validation population was identical to the comparison analysis population (members aged years), the QFracture population consisted of members, Sensitivity Hip fractures Comparative analysis Major osteoporotic fractures Top 10% cut-off Top 20% cut-off QFracture (50-90 years; AUC = 82.7%) FRAX (50-90 years; AUC = 81.5%) Garvan (50-90 years; AUC = 77.8%) QFracture (50-90 years; AUC = 71.2%) FRAX (50-90 years; AUC = 71.4%) Sensitivity Tool specific external validation QFracture ( years; AUC = 88.0%) 0.2 FRAX (50-90 years; AUC = 81.5%) Garvan (60-95 years; AUC = 71.2%) specificity QFracture ( years; AUC = 75.4%) FRAX (50-90 years; AUC = 71.4%) specificity Fig 2 Receiver operating curves of QFracture, FRAX, and Garvan predictive tools for hip and major osteoporotic fractures during five years of follow-up in comparative and tool specific external validation analyses 8 doi: /bmj.i6755 BMJ 2017;356:i6755 the bmj

9 Table 3 Reclassification analysis* for QFracture compared with FRAX Hip fractures Major osteoporotic fractures QFracture FRAX Reclassification measures QFracture FRAX Reclassification measure High risk, (10% of population) Low risk, (90% of population) (total population) High risk, (10% of population) Low risk, (90% of population) (total population) NRI-ne (SE) 0.19% (0.03%) the bmj BMJ 2017;356:i6755 doi: /bmj.i (II) (correctly reclassified) (I) (correctly classified by both models) 0.04% (0.03%) Low risk, (90% of population) NRI-ne (SE) (II) (correctly reclassified) (I) (correctly classified by both models) NRI-e (SE) 2.31% (0.14%) Low risk, (90% of population) 7934 (IV) (incorrectly reclassified) (VI) (correctly classified by both models) (VIII) (misclassified by both models) NRI-e (SE) 1.50% (0.27%) (III) (misclassified by both models) NRI (SE) 1.54% (0.27%) High risk, 6051 (V) (correctly (10% of reclassified) WNRI (SE) 0.08% (0.05%) population) (VII) (incorrectly reclassified) 2651 (IV) (incorrectly reclassified) (III) (misclassified by both models) NRI (SE) 2.50% (0.15%) 9609 (VI) (correctly classified by both models); 3072 (V). (correctly reclassified) WNRI (SE) 0.36% (0.05%) High risk, (10% of population) (VIII) (misclassified by both models) (VII) (incorrectly reclassified) NRI=net reclassification index; NRI-e=net reclassification for events; NRI-ne=net reclassification for non-events; WNRI=weighted net reclassification index; SE=standard error. NRI calculated as NRI-e+NRI-ne. NRI-ne calculated as (V IV)/(III+IV+V+VI. NRI-e calculated as (II VI)/(I+II+VII+VIII). WNRI calculated as NRI-e ((III+IV+V+VI)/IX)+NRI-ne ((I+II+VII+VIII)/IX). *Calculated with five years of follow-up. High risk group defined by each model as study participants who received a risk score in top 10% of risk, and the low risk as the 90% who did not. Non-events. Events (people who sustained a fracture). aged years, and the Garvan population included members, aged years. The population of the QFracture external validation included (1.7%) individuals who experienced a hip fracture and (5.2%) individuals who experienced a major osteoporotic fracture during the follow-up period. The corresponding rates for the population of the Garvan external validation were (4.2%) and (10.3%), respectively. Supplement 8 provides a comparison of the prevalence of the risk factors between the populations used to develop the tools (derivation cohorts), and the population of the tool specific external validations in the current study for QFracture 16 and FRAX. 38 The prevalence of risk factors as defined in the final Garvan model were not available for the original Garvan population The current study s QFracture tool specific population was relatively older than QFracture s derivation cohort and was characterised by a greater prevalence (or greater capture rates) of most risk factors. In contrast, the current study s FRAX tool specific population was similar in age to FRAX s derivation cohort, with a smaller share of women and lower prevalence (or lower capture rates) of risk factors. AUC values for hip fracture in the validation analyses were 88.0% (95% confidence interval 87.8% to 88.2%) for QFracture, 81.5% (81.3% to 81.7%) for FRAX, and 71.2% (70.9% to 71.5%) for Garvan (table 6 ). The Garvan hip fracture tool was the only one to present sex specific AUC and sensitivity values that were both higher than the overall values. Figure 2 presents the comparisons of the receiver operating characteristic curves for the tool specific external validations. Supplement 9 provides calibration analyses for age and 10ths of risk groups for each of the tool specific external validation cohorts. Discussion This study included over one million adults aged in a single, general population and directly compared the three most studied fracture prediction tools in an electronic health record system. The discriminatory performance according to the area under the receiver operating curve (AUC) of hip fracture scores for both FRAX and QFracture was high, with the latter performing slightly better, followed by a moderate performance of Garvan. Discriminatory measures for the prediction of major osteoporotic fractures were lower overall than for hip fracture prediction, with very close AUC measures for FRAX and QFracture. Three different sensitivity analyses (see supplements 5-7) examining the impact of input data definitions as well as a different population definition among patients naïve to osteoporosis treatment, have all supported these findings. Given that small differences in the overall AUC (as observed between QFracture and FRAX) may not reflect the entire difference in the discriminative performance for individual patients with a unique set of risk factors, we evaluated the reclassification of individuals between these tools. In examining the value gained from the additional risk factors included in QFracture compared with FRAX, reclassification analysis showed that QFracture had an overall 0.08% net increase and a 9

10 Table 4 Calibration* of observed versus predicted hip fracture rates, by sex and age groups Women Men Age range No of people Hip fracture rate (%) (No with first hip fracture) Mean (SD) predicted probability (%) Observed to predicted ratio No of people Hip fracture rate (%) (No with first hip fracture) Mean (SD) predicted probability (%) QFracture: (389) 0.1 (0.002) (415) 0.1 (0.007) (657) 0.2 (0.004) (533) 0.2 (0.005) (954) 0.4 (0.004) (689) 0.3 (0.007) (1284) 0.9 (0.012) (762) 0.6 (0.015) (2520) 1.9 (0.023) (1250) 1.2 (0.026) (4025) 3.9 (0.050) (1670) 2.6 (0.052) (5241) 7.5 (0.087) (2044) 5.3 (0.088) (4025) 11.9 (0.122) (1633) 9.6 (0.133) 0.9 FRAX: (389) 0.2 (0.002) (415) 0.1 (0.001) (657) 0.4 (0.003) (533) 0.3 (0.002) (954) 0.7 (0.005) (689) 0.4 (0.003) (1284) 1.2 (0.010) (762) 0.8 (0.006) (2520) 2.4 (0.019) (1250) 1.5 (0.011) (4025) 4.3 (0.030) (1670) 2.5 (0.017) (5241) 6.0 (0.036) (2044) 3.4 (0.019) (4025) 6.8 (0.037) (1633) 3.8 (0.020) 2.3 Garvan: (389) 0.1 (0.001) (415) 0.1 (0.001) (657) 0.1 (0.002) (533) 0.2 (0.002) (954) 0.2 (0.004) (689) 0.4 (0.004) (1284) 0.4 (0.009) (762) 0.8 (0.008) (2520) 0.6 (0.016) (1250) 1.6 (0.016) (4025) 1.3 (0.033) (1670) 3.5 (0.033) (5241) 2.6 (0.057) (2044) 7.1 (0.065) (4025) 4.7 (0.093) (1633) 13.9 (0.109) 0.6 *Assessed with five years of follow-up. Observed to predicted ratio 0.36% net decrease in the proportion of patients assigned a more appropriate risk category for hip fractures and major osteoporotic fractures, respectively. The combination of these results suggests an overall similar discriminatory performance for QFracture and FRAX, with a small advantage in hip fracture prediction for the former and a small advantage in major osteoporotic fracture prediction for the latter. The tool specific external validation analyses presented comparable results to those reported in previous individual tool validations of the same age ranges. Despite the identical age ranges that were used for the tool specific external validations, the populations still differed to some extent from the derivation cohorts to which they were compared in terms of overall average age, sex distribution, and prevalence of risk factors (see supplement 8). In addition, the FRAX derivation cohort excluded patients treated for osteoporosis, but the current study found very similar results for FRAX when tested in a cohort with and without these patients (see supplement 7). Owing to these differences, our tool specific external validation analyses provided evidence for the transportability of the tools when considering the spectrum of external validation studies ranging from reproducible to transportable. Furthermore, by comparing performance gaps between tools both in the same population and in populations of different age ranges, our analyses substantiated previous claims of a strong correlation between age spans of the studied population and the observed performance of the tested tool In an analysis of the calibration measures, FRAX presented the best observed-to-predicted ratios, with the weighted average closest to 1, both across age groups and across predicted risk 10ths. Additionally, the calibration slopes of FRAX were closest to 1, representing better calibration across individuals, on top of the better calibration among groups. The FRAX calibration ratios were also relatively stable, whereas QFracture and Garvan presented a decline in the observed-to-predicted ratios as age increased. A possible contributor to FRAX s more consistent observed-to-predicted ratio across age groups is that it accounts for the competing risk of death, whereas Garvan and QFracture do not. 4 The integration of competing death risk into fracture prediction simulates real world behaviour by assigning lower predicted fracture rates for groups of individuals who have lower life expectancy, such as older people. The issue of whether competing risk of death should be incorporated into fracture prediction tools has been debated in the literature, with some studies accounting for it and others not Our comparative results observed within a single real world population illustrate that calibration is relatively more consistent when competing risk is incorporated. The observed-to-predicted ratios of QFracture and Garvan also presented a declining trend over 10ths of risk. The trend observed over 10ths is at least in part likely explained by age, 10 doi: /bmj.i6755 BMJ 2017;356:i6755 the bmj